Leveraging Machine Learning for Bidding Strategies in Miner Extractable Value Auctions
TLDR
This paper analyzes changes in MEV bidding/auction behaviors. It shows that constant vs dynamic bidding strategies are optimal for different MEV attack vector. The paper proposes an ML model for bidding that wins over 50% of all Flashbot auctions.
Key learnings
The application of MEV strategies is not monolithic. That is to say, the type of MEV in question dictates the optimal strategies utilized by various participants.
ML strategies are well-suited to auction/bidding mechanisms and are easily optimizable and dynamic according to changes in participant and market behavior.
The ML strategy proposed in this paper is valuable and applicable to retail vault use cases for Concrete.
Concrete
Once retail vaults are implemented for Concrete coverage uses, bidding mechanics are highly relevant in order to optimize yield and liquidity curves.
If Concrete is successful, these auctions will be as competitive as MEV opportunities and structuring mechanisms to account for this competitiveness is vital prior to launch.
The ML model provided is generalizable and applicable to retail vault use cases for Concrete.
Details
Unlike P2P network broadcasting, MEV extractors using a relay service cannot observe their competitors’ bids during the auction, leading to the adoption of sealed-bidding strategies.
The closed/sealed nature of these auctions has been greatly exacerbated by Flashbot private pools.
MEV competitions are first-price sealed-bid auctions, which allows MEV extractors to privately communicate their bid and transaction order preference without paying for failed bids.
The auction mechanism tries to increase validator payoffs, while non-winning participants remain anonymous due to the sealed auction format.
The auction also provides guarantees such as pre-trade privacy and failed trade privacy. Bribes can either be paid through direct payment to the validator or higher gas fees.
Top 10 bots won ~50% of all auctions. This is extremely high.
MEV opportunities are categorized as sandwich attacks or cyclical arbitrages
Sandwich attacks are well defined - placing a transaction between two others to swap and profit on slippage.
2 kinds of cyclical arbitrages which amount to the same results - profit based on swap differences.
Bribe ratios for both MEV options have approached, and stayed near, 100% value of profit.
ML model was trained on the following features:
Block number
Potential profit
Revenue with base fee
Number of swaps
Start amount
End amount
Base gas cost
Gas required
Protocols
Token categories
Resulting model was built on on an LGBM Regressor model. This model won 50%+ of all profit values, not just auctions.
Model shows that dynamic vs static models are not clearly hierarchical regarding value/profit. For example, sandwich attacks are better-suited for dynamic/adaptive bidding strategies.
This situation is analogous to how a retail vault may be structured for us. It is worth understanding models that we, or others, could use in participating with them.
Challenges/concerns/comments
This paper is of the highest quality - the methods are beyond reproach. No concerns.
The model utilized may not be easily adaptable to other auction use cases given its selection based on relevant features.
Further reading
Original paper link
Last updated